[Bug d/106977] [13 regression] d21 dies with SIGBUS on 32-bit Darwin

["ibuclaw at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106977

--- Comment #20 from ibuclaw at gcc dot gnu.org ---
(In reply to Andrew Pinski from comment #19)
> (In reply to Andrew Pinski from comment #18)
> > > I think the visibility type is POD (assuming D has that concept)
> 
> At least the front-end does.
> See dmd/dstruct.d:443
> 
> if isPOD return false, TREE_ADDRESSABLE is set on the struct.
> I have not gone through the code otherwise.
See d/decl.cc:950

It's not TREE_ADDRESSABLE, but on 32-bit the struct is considered to be
'aggregate_value_p', which in turn sets up return by hidden reference.

This effects how the return is handled later (around toir.cc:1044), which
splits up the init and return expression.  Returning this way I guess is fine
for extern(D) functions, however we should not be so eager to do rvo/sret for
other extern language functions.

Having a quick look at C++ front-end, they require both `aggregate_value_p` and
for a named variable to be in the return expression.

```
  if (current_function_return_value)
    {
      tree r = current_function_return_value;
      tree outer;

      if (r != error_mark_node
          /* This is only worth doing for fns that return in memory--and
             simpler, since we don't have to worry about promoted modes.  */
          && aggregate_value_p (TREE_TYPE (TREE_TYPE (fndecl)), fndecl)
```

So for the current gate in D:
```
    if (TREE_ADDRESSABLE (TREE_TYPE (resdecl))
        || aggregate_value_p (TREE_TYPE (resdecl), fndecl))
```

I think tightening that to (its late, and my parentheses may be wrong).
```
    if (TREE_ADDRESSABLE (TREE_TYPE (resdecl))
        || ((d->resolvedLinkage () == LINK::d
             || (d->resolvedLinkage () == LINK::cpp && d->nrvo_var))
            && aggregate_value_p (TREE_TYPE (resdecl), fndecl)))
```
Which is:
1. TREE_ADDRESSABLE
2. extern(D) and aggregate_value_p
3. extern(C++) and have NVRO variable and aggregate_value_p

I guess for extern(C) functions we should just forget even attempting to do any
(N)RVO/SRET returns and let tree-nrv.cc decide whether to optimize or not.